AI product notes: week of 20 July 2026

Smaller product updates do not always need a full article, but they can still matter if you own the platform, run the admin console, or build against the API.

Updates

• Anthropic: Claude Code /doctor command audits your setup for unused skills, bloated CLAUDE.md, and slow hooks (26 July 2026). As Claude Code projects accumulate plugins, skills, and CLAUDE.md cruft, /doctor gives teams a systematic way to control context costs and maintain performance without manual archaeology.

• OpenAI: Codex expands /import to migrate Cursor and Claude Code settings, MCP servers, plugins, sessions, and project-scoped memories (22 July 2026). Seamless one-command import of an entire Claude Code or Cursor environment removes the largest friction point for developers evaluating Codex, intensifying the coding-agent switching war.

• OpenAI: Codex fixes bundled instructions and corrects 272,000-token context windows for GPT-5.6 Sol, Terra, and Luna (22 July 2026). Incorrect context-window metadata for a just-launched model family causes silent failures in any Codex pipeline that dynamically chunks inputs based on reported token limits.

• Google: Gemini Enterprise Agent Platform patches an SSRF vulnerability in Agent Studio web app code (22 July 2026). An SSRF flaw in the agent-building surface of a widely deployed enterprise AI platform is a governance-relevant finding, even though Google patched it server-side without requiring customer action.

• Google: Gemini Omni comes to Google Vids for higher-quality video generation and text-based edits (22 July 2026). Gemini Omni’s arrival in Google Vids makes AI video generation and editing a native Workspace capability for the first time, relevant to any team already using Google Workspace for content creation.

• Microsoft: Microsoft Agent Framework ships 1.12.0 with Azure Cosmos DB semantic memory, MCP/A2A hosting helpers, and a major .NET update (22 July 2026). Stable MCP and A2A hosting support in the official Microsoft agent SDK is a meaningful step for enterprise teams building cross-vendor agent workflows on Azure.

• Microsoft: Microsoft MCAPS Start for Partners FY27 kickoff frames Copilot and agents as the core execution priority (22 July 2026). Microsoft is formally reorienting its partner channel around Copilot production deployments for FY27, retiring existing certifications and introducing Agentic AI-focused credentials that resellers and consultants must now pursue.

• OpenAI: OpenAI’s safety-alignment blog post on long-horizon models coincides with the Hugging Face incident disclosure (22 July 2026). The safety post and the incident together reveal OpenAI’s current thinking on agentic risk and the controls it intends to apply — important context for any organisation evaluating OpenAI models in autonomous workflows.

• OpenAI: ChatGPT desktop app ships unified Chat/Work layout with cloud-synced Work conversations across all devices (21 July 2026). The desktop refresh finalises the merger of ChatGPT chat and agentic Work into a single surface and adds cross-device continuity — a usability upgrade that will affect every daily ChatGPT desktop user.

• Google: Gemini 3.5 Flash-Lite rolls out in the Gemini app and Google Search (21 July 2026). Flash-Lite coming to Google Search means hundreds of millions of users may interact with it in AI Mode, making it one of the widest-reach model deployments Google has made.

• Anthropic: Claude Code changes /verify and /code-review to manual-only — they no longer run autonomously (20 July 2026). Any workflow or CLAUDE.md configuration that relied on Claude automatically triggering code review will silently stop doing so — teams need to audit their setups.

• Google: Gemini Enterprise’s Idea Generation agent is removed — users should switch to the main assistant or Deep Research (20 July 2026). Customers and admins who built workflows or user guidance around the dedicated Idea Generation agent need to update those before users hit a missing feature.

• Google: Google reportedly building ‘Frozen v2’ chip to run Gemini 6–10× more efficiently — but no official confirmation (20 July 2026). If accurate, a chip optimized for Gemini at this efficiency level would materially reduce Google’s inference cost curve and ease the compute capacity constraints that have already affected enterprise API customers.

• Microsoft: M365 Copilot gets tenant-wide Prompt Gallery with admin-published prompt collections (20 July 2026). Org-wide prompt standardization reduces AI adoption friction and ensures Copilot responses align with internal terminology and processes — a meaningful step for enterprise governance.

• Anthropic: Mid-conversation system messages now available on Claude Fable 5, Mythos 5, and Opus 4.8 — no beta header required (20 July 2026). Mid-conversation system message injection is a key building block for dynamic agent instructions; GA availability with no header removes a friction point for production deployments.